Effect of Sliding Window Sizes on Sensor-Based Human Activity Recognition Using Smartwatch Sensors and Deep Learning Approaches

Smartwatch sensors for human activity recognition (HAR) have gained significant attention due to their applications in healthcare and fitness monitoring. The effectiveness of HAR systems largely depends on the choice of sliding window widths for sensor data segmentation. This study investigates the impact of varying sliding window widths on the accuracy of HAR using wristwatch sensors and deep learning techniques. We conducted experiments using the daily human activity (DHA) dataset, comprising sensor data from 11 distinct activities. Data was preprocessed and segmented using window sizes ranging from 5 to 40 seconds. Four deep learning models (CNN, LSTM, BiLSTM, and CNN-LSTM) were employed and evaluated using accuracy, precision, recall, and F1-score. Window size significantly affected HAR performance. Smaller windows improved short-duration activity recognition but increased computational complexity, while larger windows reduced computational load but decreased accuracy for rapid activity changes. The CNN-LSTM hybrid model consistently outperformed other models, achieving 92.11% accuracy with a 20-second window and overlapping segmentation. This research provides valuable insights into balancing recognition accuracy and computational resources in smartwatch sensor-based HAR, contributing to the development of efficient and accurate systems for real-world applications.